Inconsistency-Aware Minimization: Improving Generalization with Unlabeled Data

Authors: Hee-Sung Kim, Hyeonsung Kim, Sungyoon Lee

Venue: International Conference on Machine Learning (ICML), 2026

We propose Inconsistency-Aware Minimization (IAM), which recovers SAM’s flat-minima bias from unlabeled data alone by regularizing local inconsistency — an output-sensitivity measure tied to the largest eigenvalue of the Fisher Information Matrix. IAM matches SAM in supervised learning while uniquely leveraging unlabeled data to improve semi- and self-supervised methods such as FixMatch and SimCLR.

Download Paper | arXiv | Poster | Slides | BibTeX